In [44]:
temp = cv.imread(normal_samples[0])
plt.imshow(temp)
Out[44]:
<matplotlib.image.AxesImage at 0x7cd65d462150>
# This Python 3 environment comes with many helpful analytics libraries installed
# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python
# For example, here's several helpful packages to load
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
import cv2 as cv
import matplotlib.pyplot as plt
# Input data files are available in the read-only "../input/" directory
# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory
ims = []
import os
for dirname, _, filenames in os.walk('/kaggle/input'):
for filename in filenames:
ims.append(os.path.join(dirname, filename))
# print(ims)
# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using "Save & Run All"
# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session
normal_samples = []
pneumonia_samples = []
for im in ims:
if "NORMAL" in im:
normal_samples.append(im)
elif "PNEUMONIA" in im:
pneumonia_samples.append(im)
print("Normal samples:", len(normal_samples))
print("Pneuomonia samples:", len(pneumonia_samples))
temp = cv.imread(normal_samples[0])
plt.imshow(temp)
<matplotlib.image.AxesImage at 0x7cd65d462150>
temp = cv.imread(pneumonia_samples[0])
plt.imshow(temp)
<matplotlib.image.AxesImage at 0x71f0651528d0>
def preprocessImage(imgs_inp):
preprocessed = []
for im in imgs_inp:
im = cv.imread(im)
if type(im) == np.ndarray:
im = cv.cvtColor(im, cv.COLOR_BGR2GRAY)
im = cv.blur(im, (3,3))
preprocessed.append(cv.resize(im, (32, 32)))
return preprocessed
normalpre = preprocessImage(normal_samples)
pneumoniapre = preprocessImage(pneumonia_samples)
n_train = normalpre[:int((80/100)*len(normalpre))]
p_train = pneumoniapre[:int((80/100)*len(pneumoniapre))]
y_train = [0]*len(n_train) + [1]*len(p_train)
n_test = normalpre[int((80/100)*len(normalpre)):]
p_test = pneumoniapre[int((80/100)*len(pneumoniapre)):]
y_test = [0]*len(n_test) + [1]*len(p_test)
n_train, p_train, n_test, p_test = np.array(n_train)/255, np.array(p_train)/255, np.array(n_test)/255, np.array(p_test)/255
x_train = np.array(list(n_train) + list(p_train))
x_test = np.array(list(n_test) + list(p_test))
print(len(x_train), len(y_train), len(x_test), len(y_test))
# print(x_train)
9368 9368 2344 2344
x_train = x_train.reshape((x_train.shape[0], x_train.shape[1], x_train.shape[2], 1))
x_test = x_test.reshape((x_test.shape[0], x_test.shape[1], x_test.shape[2], 1))
y_train = np.array(y_train)
y_test = np.array(y_test)
x_train.shape, x_test.shape, y_train.shape, y_test.shape
((9368, 32, 32, 1), (2344, 32, 32, 1), (9368,), (2344,))
from tensorflow import keras
from keras.layers import Conv2D, MaxPool2D, Flatten, Dense
model = keras.Sequential([
Conv2D(32, (3,3), activation="relu", input_shape = (32, 32, 1)),
MaxPool2D(pool_size=(2,2)),
Conv2D(64, (3,3), activation="relu"),
MaxPool2D(pool_size=(2,2)),
Flatten(),
Dense(1024, activation="relu"),
Dense(200, activation="relu"),
Dense(2, activation="softmax")
])
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=['accuracy'])
model.fit(x_train, y_train, epochs=20)
Epoch 1/20 293/293 [==============================] - 2s 4ms/step - loss: 0.2446 - accuracy: 0.8982 Epoch 2/20 293/293 [==============================] - 1s 3ms/step - loss: 0.1236 - accuracy: 0.9539 Epoch 3/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0998 - accuracy: 0.9632 Epoch 4/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0777 - accuracy: 0.9713 Epoch 5/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0631 - accuracy: 0.9767 Epoch 6/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0490 - accuracy: 0.9822 Epoch 7/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0467 - accuracy: 0.9836 Epoch 8/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0378 - accuracy: 0.9869 Epoch 9/20 293/293 [==============================] - 1s 5ms/step - loss: 0.0343 - accuracy: 0.9877 Epoch 10/20 293/293 [==============================] - 1s 4ms/step - loss: 0.0292 - accuracy: 0.9896 Epoch 11/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0264 - accuracy: 0.9900 Epoch 12/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0138 - accuracy: 0.9950 Epoch 13/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0117 - accuracy: 0.9963 Epoch 14/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0254 - accuracy: 0.9901 Epoch 15/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0194 - accuracy: 0.9923 Epoch 16/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0201 - accuracy: 0.9930 Epoch 17/20 293/293 [==============================] - 1s 4ms/step - loss: 0.0095 - accuracy: 0.9966 Epoch 18/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0177 - accuracy: 0.9934 Epoch 19/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0165 - accuracy: 0.9942 Epoch 20/20 293/293 [==============================] - 1s 3ms/step - loss: 0.0104 - accuracy: 0.9959
<keras.callbacks.History at 0x7cd5c14e0510>
model.evaluate(x_test, y_test)
74/74 [==============================] - 0s 3ms/step - loss: 0.0104 - accuracy: 0.9966
[0.010409509763121605, 0.9965870380401611]
As seen above, the accuracy on the original dataset is approximately 99.66%.
import sys, os, glob, time, imageio
import numpy as np, pandas as pd
import matplotlib.pyplot as plt
import matplotlib.animation as animation
from IPython.display import HTML
from PIL import Image
import torch
import torchvision.utils as vutils
import torchvision.transforms as transforms
from keras import models, layers, optimizers
from keras.models import Sequential
from keras.preprocessing.image import array_to_img, img_to_array, load_img
import tensorflow as tf
# Root directory for dataset
path_root = '../input/chest-xray-pneumonia/chest_xray/chest_xray/'
path_train= '../input/chest-xray-pneumonia/chest_xray/chest_xray/train/'
path_test = '../input/chest-xray-pneumonia/chest_xray/chest_xray/test/'
path_val = '../input/chest-xray-pneumonia/chest_xray/chest_xray/val/'
# Root paths for X-Ray images
XRay_normal = glob.glob(path_root+'*/NORMAL/*.jpeg', recursive=True)
XRay_pneumonial = glob.glob(path_root+'*/PNEUMONIA/*.jpeg', recursive=True)
# Root paths for X-Ray training images
train_normal = glob.glob(path_train+'NORMAL/*.jpeg', recursive=True)
train_pneumonial= glob.glob(path_train+'PNEUMONIA/*.jpeg', recursive=True)
# X-Ray testing images
test_normal = os.listdir(path_test+'NORMAL/')
test_pneumonial = os.listdir(path_test+'PNEUMONIA/')
print('The Chest X-Ray Images dataset contains {:04d} NORMAL and {:04d} PNEUMONIA images and {:03d} images in total'\
.format(len(XRay_normal),
len(XRay_pneumonial),
len(glob.glob(path_root+'*/*/*.jpeg')),))
print('The training dataset has {:04d} NORMAL, {:04d} PNEUMONIA and {:04d} images total'\
.format(len(train_normal),
len(train_pneumonial),
len(glob.glob(path_train+'*/*.jpeg'))))
print('The testing dataset has {:03d} NORMAL and {:03d} PNEUMONIA and {:03d} images total'\
.format(len(test_normal),
len(test_pneumonial),
len(glob.glob(path_test+'*/*.jpeg'))))
print('The validation dataset has {:01d} NORMAL and {:01d} PNEUMONIA and {:02d} images total'\
.format(len(glob.glob(path_val+'NORMAL/*.jpeg')),
len(glob.glob(path_val+'PNEUMONIA/*.jpeg')),
len(glob.glob(path_val+'*/*.jpeg'))))
The Chest X-Ray Images dataset contains 1583 NORMAL and 4273 PNEUMONIA images and 5856 images in total The training dataset has 1341 NORMAL, 3875 PNEUMONIA and 5216 images total The testing dataset has 234 NORMAL and 390 PNEUMONIA and 624 images total The validation dataset has 8 NORMAL and 8 PNEUMONIA and 16 images total
nrows, ncols = 4, 7
plt.figure(figsize=(16,10))
for idx, name in enumerate(test_normal[:nrows*ncols]):
plt.subplot(nrows, ncols, idx+1)
img = Image.open(path_test+'NORMAL/'+name)
img = img.resize(size=(128, 128), resample=Image.ANTIALIAS, box=None)
plt.imshow(img)
plt.title(name[:-5], fontsize=9)
plt.axis('off')
plt.show()
/opt/conda/lib/python3.7/site-packages/ipykernel_launcher.py:6: DeprecationWarning: ANTIALIAS is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.LANCZOS instead.
def get_data(data_path, dim=(128, 128), rand_shuffle=True):
start = time.time()
imgs_data = []
sample_size = len(data_path)
for idx, im_path in enumerate(data_path):
if idx%(sample_size//10)==0:
print('Processing index {:05d} of {:05d} ==> {:03d}%'\
.format(idx, sample_size, round(100*idx/sample_size)))
img = img_to_array(load_img(im_path, target_size = dim))
imgs_data.append(img)
# to float
imgs_data = np.array(imgs_data).astype('float32')
# scale to [0,1] (note the . after 255 - float)
imgs_data = imgs_data/255.
# shuffle the data
if rand_shuffle:
idx = np.arange(imgs_data.shape[0])
np.random.shuffle(idx)
imgs_data = imgs_data[idx,:,:,:]
return imgs_data
print('Starting for NORMAL X-Ray images ...')
X_normal = get_data(XRay_normal)
print()
print('Starting for PNEUMONIA X-Ray images ...')
X_pneumonial = get_data(XRay_pneumonial)
Starting for NORMAL X-Ray images ... Processing index 00000 of 01583 ==> 000% Processing index 00158 of 01583 ==> 010% Processing index 00316 of 01583 ==> 020% Processing index 00474 of 01583 ==> 030% Processing index 00632 of 01583 ==> 040% Processing index 00790 of 01583 ==> 050% Processing index 00948 of 01583 ==> 060% Processing index 01106 of 01583 ==> 070% Processing index 01264 of 01583 ==> 080% Processing index 01422 of 01583 ==> 090% Processing index 01580 of 01583 ==> 100% Starting for PNEUMONIA X-Ray images ... Processing index 00000 of 04273 ==> 000% Processing index 00427 of 04273 ==> 010% Processing index 00854 of 04273 ==> 020% Processing index 01281 of 04273 ==> 030% Processing index 01708 of 04273 ==> 040% Processing index 02135 of 04273 ==> 050% Processing index 02562 of 04273 ==> 060% Processing index 02989 of 04273 ==> 070% Processing index 03416 of 04273 ==> 080% Processing index 03843 of 04273 ==> 090% Processing index 04270 of 04273 ==> 100%
def _time(start, end):
# if in seconds
if (end-start)<60:
wall_time = f'{round((end-start),2)}sec'
# if in minute(s)
elif (end-start)>=3600:
wall_time = f'{int((end-start)/3600)}h {int(((end-start)%3600)/60)}min {round((end-start)%60,2)}sec'
# if in houre(s)
else:
wall_time = f'{int((end-start)/60)}min {round((end-start)%60,2)}sec'
return wall_time
def define_grid(data_images, nrows=4, ncols=5, plot_grid=True):
start = time.time()
ngpu = 1
device = torch.device("cuda:0" if (torch.cuda.is_available() and ngpu > 0) else "cpu")
# Rearange the shape of the data
data_transp = [np.transpose(data_images[i,:,:]) for i in range(data_images[:nrows*ncols].shape[0])]
# From to torch type for the grid
data_transp = torch.Tensor(data_transp)
print(f'The shape is reordered from {data_images.shape[1:]} to {data_transp.shape[1:]} in {_time(start, time.time())}')
# Make the grid
grid_images = np.transpose(
vutils.make_grid(
data_transp.to(device)[:nrows*ncols],
nrow=nrows,
padding=2,
normalize=True,
scale_each=True,
pad_value=1,
).cpu(), axes=(2,1,0))
# Show the output grid
if plot_grid:
plt.figure(figsize=(12,12))
plt.axis("off")
plt.title(f'Grid of {nrows*ncols} real images', fontsize=27)
plt.imshow(grid_images)
return grid_images
grid_X_normal = define_grid(X_normal, plot_grid=False)
grid_X_pneumonial = define_grid(X_pneumonial, plot_grid=False)
/opt/conda/lib/python3.7/site-packages/ipykernel_launcher.py:8: UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor. (Triggered internally at /usr/local/src/pytorch/torch/csrc/utils/tensor_new.cpp:207.)
The shape is reordered from (128, 128, 3) to torch.Size([3, 128, 128]) in 0.19sec The shape is reordered from (128, 128, 3) to torch.Size([3, 128, 128]) in 0.08sec
fig, (ax1, ax2)= plt.subplots(nrows = 1, ncols = 2, figsize = (19, 8))
ax1.imshow(grid_X_normal); ax1.axis('off')
ax1.set_title(label = 'Grid of X-Ray NORMAL images', fontsize = 27)
ax2.imshow(grid_X_pneumonial); ax2.axis('off')
ax2.set_title(label = 'Grid of X-Ray PNEUMONIA images', fontsize = 27)
plt.tight_layout(pad = 1.08, h_pad = None, w_pad = None, rect = [0, 0.03, 1, 0.95])
n_epoch = 200
batch_size = 128
latent_dim = 100
cols, rows = 128, 128
channels = 3
dim = cols, rows
in_shape = (cols, rows, channels)
lr = 0.0002
beta1 = 0.5
ngpu = 1
nrows, ncols = 3, 4
def define_discriminator(in_shape = (128,128,3)):
model = models.Sequential()
# normal
model.add(layers.Conv2D(128, (5,5), padding = 'same', input_shape = in_shape))
model.add(layers.LeakyReLU(alpha = 0.2))
# downsample to 64x64
model.add(layers.Conv2D(128, (5,5), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# downsample to 32x32
model.add(layers.Conv2D(128, (5,5), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# downsample to 16x16
model.add(layers.Conv2D(128, (5,5), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# downsample to 8x8
model.add(layers.Conv2D(128, (5,5), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# classifier
model.add(layers.Flatten())
model.add(layers.Dropout(0.4))
model.add(layers.Dense(1, activation = 'sigmoid'))
# compile model
opt = tf.keras.optimizers.Adam(lr = 0.0002, beta_1 = 0.5)
model.compile(loss = 'binary_crossentropy', optimizer = opt, metrics = ['accuracy'])
return model
def define_generator(latent_dim):
model = models.Sequential()
# foundation for 8x8 feature maps
n_nodes = 128*8*8
model.add(layers.Dense(n_nodes, input_dim = latent_dim))
model.add(layers.LeakyReLU(alpha = 0.2))
model.add(layers.Reshape((8, 8, 128)))
# upsample to 16x16
model.add(layers.Conv2DTranspose(128, (4,4), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# upsample to 32x32
model.add(layers.Conv2DTranspose(128, (4,4), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# upsample to 64x64
model.add(layers.Conv2DTranspose(128, (4,4), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# upsample to 128x128
model.add(layers.Conv2DTranspose(128, (4,4), strides = (2,2), padding = 'same'))
model.add(layers.LeakyReLU(alpha = 0.2))
# output layer 128x128x3
model.add(layers.Conv2D(3, (5,5), activation = 'tanh', padding = 'same'))
return model
def generate_latent_points(latent_dim, n_samples):
x_input = np.random.randn(latent_dim*n_samples)
x_input = x_input.reshape(n_samples, latent_dim)
return x_input
# use the generator to generate n fake examples, with class labels
def generate_fake_samples(g_model, latent_dim, n_samples):
x_input = generate_latent_points(latent_dim, n_samples)
X = g_model.predict(x_input)
# create 'fake' class labels (0)
y = np.zeros((n_samples, 1))
return X, y
def define_gan(g_model, d_model):
d_model.trainable = False
model = models.Sequential()
model.add(g_model)
model.add(d_model)
opt = tf.keras.optimizers.Adam(lr = 0.0002, beta_1 = 0.5)
model.compile(loss = 'binary_crossentropy', optimizer = opt)
return model
# retrive real samples
def get_real_samples(dataset, n_samples):
ix = np.random.randint(0, dataset.shape[0], n_samples)
X = dataset[ix]
y = np.ones((n_samples, 1))
return X, y
# create and save a plot of generated images
def show_generated(generated, epoch, nrows=4, ncols=5):
plt.figure(figsize=(10,10))
for idx in range(nrows*ncols):
plt.subplot(nrows, ncols, idx+1)
plt.imshow(generated[idx])
plt.axis('off')
plt.savefig('image_at_epoch_{:04d}.png'.format(epoch+1))
plt.show()
# evaluate the discriminator and plot generated images
def summarize_performance(epoch, g_model, d_model, dataset, latent_dim, n_samples=100):
# prepare real samples
X_real, y_real = get_real_samples(dataset, n_samples)
# evaluate discriminator on real examples
_, acc_real = d_model.evaluate(X_real, y_real, verbose=0)
# prepare fake examples
x_fake, y_fake = generate_fake_samples(g_model, latent_dim, n_samples)
# evaluate discriminator on fake examples
_, acc_fake = d_model.evaluate(x_fake, y_fake, verbose=0)
# summarize discriminator performance
print('> Accuracy at epoch %d [real: %.0f%%, fake: %.0f%%]'%(epoch+1, acc_real*100, acc_fake*100))
show_generated(x_fake, epoch)
def plot_loss(loss):
plt.figure(figsize=(10,5))
plt.title("Generator and Discriminator Loss During Training", fontsize=20)
plt.plot(loss[0], label="D_real")
plt.plot(loss[1], label="D_fake")
plt.plot(loss[2], label="G")
plt.xlabel("Iteration", fontsize=20); plt.ylabel("Loss", fontsize=20)
plt.legend(); plt.show()
def train(g_model, d_model, gan_model, dataset, latent_dim = 100, n_epochs = 100, n_batch = 128):
start = time.time()
bat_per_epo = int(dataset.shape[0]/n_batch)
half_batch = int(n_batch/2)
loss1, loss2, loss3 = [], [], []
fake_liste = []
print('Training Start...')
for i in range(n_epochs):
start1 = time.time()
for j in range(bat_per_epo):
# get randomly selected 'real' samples
X_real, y_real = get_real_samples(dataset, half_batch)
# update discriminator model weights
d_loss1, _ = d_model.train_on_batch(X_real, y_real)
# generate 'fake' examples
X_fake, y_fake = generate_fake_samples(g_model, latent_dim, half_batch)
# update discriminator model weights
d_loss2, _ = d_model.train_on_batch(X_fake, y_fake)
# prepare points in latent space as input for the generator
X_gan = generate_latent_points(latent_dim, n_batch)
# create inverted labels for the fake samples
y_gan = np.ones((n_batch, 1))
# update the generator via the discriminator's error
g_loss = gan_model.train_on_batch(X_gan, y_gan)
# summarize loss on this batch
loss1.append(d_loss1); loss2.append(d_loss2); loss3.append(g_loss)
print('Epoch: {:03d}/{:03d}, Loss: [D_real = {:2.3f}, D_fake = {:2.3f}, G = {:2.3f}], time: {:s}'\
.format(i+1,n_epochs,d_loss1,d_loss2,g_loss, _time(start1,time.time())))
# evaluate the model performance
if (i+1)%(n_epochs//10) == 0:
# Save and show generated images
summarize_performance(i, g_model, d_model, dataset, latent_dim)
print('Total time for training {} epochs is {} sec'.format(n_epochs, _time(start, time.time())))
# Show loss curves
loss = (loss1, loss2, loss3)
plot_loss(loss)
discriminator = define_discriminator()
generator = define_generator(latent_dim)
gan = define_gan(generator, discriminator)
/opt/conda/lib/python3.7/site-packages/keras/optimizer_v2/optimizer_v2.py:356: UserWarning: The `lr` argument is deprecated, use `learning_rate` instead. "The `lr` argument is deprecated, use `learning_rate` instead.")
train(generator, discriminator, gan, X_normal, latent_dim, n_epochs = n_epoch, n_batch = batch_size)
Training Start... Epoch: 001/200, Loss: [D_real = 0.261, D_fake = 0.177, G = 2.982], time: 23.02sec Epoch: 002/200, Loss: [D_real = 0.229, D_fake = 0.158, G = 3.870], time: 21.91sec Epoch: 003/200, Loss: [D_real = 0.343, D_fake = 0.212, G = 3.537], time: 21.26sec Epoch: 004/200, Loss: [D_real = 0.302, D_fake = 0.214, G = 3.288], time: 21.51sec Epoch: 005/200, Loss: [D_real = 0.385, D_fake = 0.155, G = 3.420], time: 22.36sec Epoch: 006/200, Loss: [D_real = 0.722, D_fake = 0.597, G = 2.801], time: 21.89sec Epoch: 007/200, Loss: [D_real = 0.220, D_fake = 0.182, G = 3.659], time: 21.56sec Epoch: 008/200, Loss: [D_real = 0.240, D_fake = 0.166, G = 3.711], time: 21.94sec Epoch: 009/200, Loss: [D_real = 0.878, D_fake = 0.780, G = 3.500], time: 21.69sec Epoch: 010/200, Loss: [D_real = 0.228, D_fake = 0.145, G = 3.681], time: 21.85sec Epoch: 011/200, Loss: [D_real = 0.193, D_fake = 0.234, G = 4.182], time: 21.92sec Epoch: 012/200, Loss: [D_real = 0.140, D_fake = 0.136, G = 4.157], time: 21.76sec Epoch: 013/200, Loss: [D_real = 0.121, D_fake = 0.264, G = 4.139], time: 21.76sec Epoch: 014/200, Loss: [D_real = 0.243, D_fake = 0.200, G = 5.674], time: 22.0sec Epoch: 015/200, Loss: [D_real = 0.274, D_fake = 0.259, G = 3.963], time: 21.67sec Epoch: 016/200, Loss: [D_real = 0.215, D_fake = 0.325, G = 4.151], time: 22.02sec Epoch: 017/200, Loss: [D_real = 0.252, D_fake = 0.145, G = 3.554], time: 21.75sec Epoch: 018/200, Loss: [D_real = 0.177, D_fake = 0.180, G = 3.511], time: 21.84sec Epoch: 019/200, Loss: [D_real = 0.292, D_fake = 0.119, G = 3.174], time: 21.75sec Epoch: 020/200, Loss: [D_real = 0.183, D_fake = 0.136, G = 3.897], time: 21.77sec > Accuracy at epoch 20 [real: 92%, fake: 100%]
Epoch: 021/200, Loss: [D_real = 0.179, D_fake = 0.241, G = 5.169], time: 21.62sec Epoch: 022/200, Loss: [D_real = 0.143, D_fake = 0.276, G = 5.047], time: 21.69sec Epoch: 023/200, Loss: [D_real = 0.181, D_fake = 0.166, G = 4.027], time: 21.69sec Epoch: 024/200, Loss: [D_real = 0.158, D_fake = 0.123, G = 3.701], time: 21.78sec Epoch: 025/200, Loss: [D_real = 0.136, D_fake = 0.133, G = 3.893], time: 21.89sec Epoch: 026/200, Loss: [D_real = 0.255, D_fake = 0.205, G = 4.743], time: 21.72sec Epoch: 027/200, Loss: [D_real = 0.303, D_fake = 0.206, G = 4.466], time: 21.66sec Epoch: 028/200, Loss: [D_real = 0.251, D_fake = 0.190, G = 3.970], time: 22.07sec Epoch: 029/200, Loss: [D_real = 0.086, D_fake = 0.127, G = 4.338], time: 21.67sec Epoch: 030/200, Loss: [D_real = 0.105, D_fake = 0.140, G = 4.410], time: 21.59sec Epoch: 031/200, Loss: [D_real = 0.147, D_fake = 0.152, G = 4.646], time: 21.75sec Epoch: 032/200, Loss: [D_real = 0.155, D_fake = 0.176, G = 4.627], time: 22.38sec Epoch: 033/200, Loss: [D_real = 0.332, D_fake = 0.125, G = 3.366], time: 21.78sec Epoch: 034/200, Loss: [D_real = 0.198, D_fake = 0.279, G = 4.500], time: 21.71sec Epoch: 035/200, Loss: [D_real = 0.058, D_fake = 0.165, G = 4.629], time: 21.69sec Epoch: 036/200, Loss: [D_real = 0.374, D_fake = 0.102, G = 3.415], time: 21.69sec Epoch: 037/200, Loss: [D_real = 0.250, D_fake = 0.241, G = 4.319], time: 21.77sec Epoch: 038/200, Loss: [D_real = 0.153, D_fake = 0.604, G = 6.519], time: 21.7sec Epoch: 039/200, Loss: [D_real = 0.228, D_fake = 0.165, G = 3.593], time: 21.69sec Epoch: 040/200, Loss: [D_real = 0.406, D_fake = 0.165, G = 3.672], time: 21.71sec > Accuracy at epoch 40 [real: 96%, fake: 99%]
Epoch: 041/200, Loss: [D_real = 0.099, D_fake = 0.134, G = 4.029], time: 21.91sec Epoch: 042/200, Loss: [D_real = 0.210, D_fake = 0.129, G = 4.234], time: 21.72sec Epoch: 043/200, Loss: [D_real = 0.136, D_fake = 0.275, G = 5.079], time: 21.57sec Epoch: 044/200, Loss: [D_real = 0.296, D_fake = 0.196, G = 4.160], time: 21.73sec Epoch: 045/200, Loss: [D_real = 0.191, D_fake = 0.291, G = 4.383], time: 21.98sec Epoch: 046/200, Loss: [D_real = 0.122, D_fake = 0.190, G = 5.015], time: 22.02sec Epoch: 047/200, Loss: [D_real = 0.167, D_fake = 0.175, G = 4.313], time: 21.96sec Epoch: 048/200, Loss: [D_real = 0.078, D_fake = 0.109, G = 4.707], time: 21.49sec Epoch: 049/200, Loss: [D_real = 0.133, D_fake = 0.091, G = 4.590], time: 21.75sec Epoch: 050/200, Loss: [D_real = 0.224, D_fake = 0.110, G = 3.876], time: 22.02sec Epoch: 051/200, Loss: [D_real = 0.965, D_fake = 0.461, G = 4.481], time: 22.13sec Epoch: 052/200, Loss: [D_real = 0.061, D_fake = 0.232, G = 5.079], time: 21.72sec Epoch: 053/200, Loss: [D_real = 0.158, D_fake = 0.249, G = 4.677], time: 21.76sec Epoch: 054/200, Loss: [D_real = 0.059, D_fake = 0.088, G = 4.555], time: 21.79sec Epoch: 055/200, Loss: [D_real = 0.504, D_fake = 0.384, G = 3.242], time: 21.62sec Epoch: 056/200, Loss: [D_real = 0.071, D_fake = 0.100, G = 3.982], time: 21.99sec Epoch: 057/200, Loss: [D_real = 0.306, D_fake = 0.222, G = 3.759], time: 21.74sec Epoch: 058/200, Loss: [D_real = 0.136, D_fake = 0.237, G = 4.292], time: 21.67sec Epoch: 059/200, Loss: [D_real = 0.990, D_fake = 0.601, G = 3.466], time: 21.78sec Epoch: 060/200, Loss: [D_real = 0.124, D_fake = 0.105, G = 4.035], time: 21.78sec > Accuracy at epoch 60 [real: 93%, fake: 100%]
Epoch: 061/200, Loss: [D_real = 0.168, D_fake = 0.127, G = 4.270], time: 21.78sec Epoch: 062/200, Loss: [D_real = 0.812, D_fake = 0.292, G = 3.984], time: 21.67sec Epoch: 063/200, Loss: [D_real = 0.093, D_fake = 0.090, G = 4.522], time: 22.06sec Epoch: 064/200, Loss: [D_real = 0.824, D_fake = 0.162, G = 3.072], time: 21.76sec Epoch: 065/200, Loss: [D_real = 0.184, D_fake = 0.101, G = 3.916], time: 21.7sec Epoch: 066/200, Loss: [D_real = 0.334, D_fake = 0.173, G = 3.164], time: 22.03sec Epoch: 067/200, Loss: [D_real = 0.169, D_fake = 0.062, G = 3.930], time: 21.67sec Epoch: 068/200, Loss: [D_real = 0.304, D_fake = 0.075, G = 3.865], time: 21.74sec Epoch: 069/200, Loss: [D_real = 0.130, D_fake = 0.107, G = 4.353], time: 21.73sec Epoch: 070/200, Loss: [D_real = 0.136, D_fake = 0.231, G = 5.187], time: 22.01sec Epoch: 071/200, Loss: [D_real = 0.139, D_fake = 0.107, G = 4.284], time: 21.76sec Epoch: 072/200, Loss: [D_real = 0.256, D_fake = 0.111, G = 4.479], time: 22.12sec Epoch: 073/200, Loss: [D_real = 0.282, D_fake = 0.149, G = 4.630], time: 21.75sec Epoch: 074/200, Loss: [D_real = 0.184, D_fake = 0.212, G = 5.085], time: 21.96sec Epoch: 075/200, Loss: [D_real = 0.655, D_fake = 0.145, G = 2.715], time: 21.52sec Epoch: 076/200, Loss: [D_real = 0.169, D_fake = 0.197, G = 5.062], time: 21.83sec Epoch: 077/200, Loss: [D_real = 0.109, D_fake = 0.095, G = 3.869], time: 22.08sec Epoch: 078/200, Loss: [D_real = 0.055, D_fake = 0.143, G = 4.707], time: 21.72sec Epoch: 079/200, Loss: [D_real = 0.233, D_fake = 0.136, G = 4.530], time: 21.65sec Epoch: 080/200, Loss: [D_real = 0.827, D_fake = 0.530, G = 5.564], time: 21.79sec > Accuracy at epoch 80 [real: 83%, fake: 100%]
Epoch: 081/200, Loss: [D_real = 0.080, D_fake = 0.113, G = 4.902], time: 21.96sec Epoch: 082/200, Loss: [D_real = 0.097, D_fake = 0.108, G = 4.350], time: 21.62sec Epoch: 083/200, Loss: [D_real = 0.065, D_fake = 0.022, G = 4.931], time: 21.64sec Epoch: 084/200, Loss: [D_real = 0.593, D_fake = 0.676, G = 4.237], time: 21.63sec Epoch: 085/200, Loss: [D_real = 0.069, D_fake = 0.144, G = 5.010], time: 21.76sec Epoch: 086/200, Loss: [D_real = 0.193, D_fake = 0.087, G = 4.156], time: 21.8sec Epoch: 087/200, Loss: [D_real = 0.209, D_fake = 0.194, G = 3.946], time: 21.79sec Epoch: 088/200, Loss: [D_real = 0.169, D_fake = 0.355, G = 5.179], time: 21.73sec Epoch: 089/200, Loss: [D_real = 0.109, D_fake = 0.212, G = 5.047], time: 21.71sec Epoch: 090/200, Loss: [D_real = 0.194, D_fake = 0.164, G = 4.860], time: 21.85sec Epoch: 091/200, Loss: [D_real = 0.085, D_fake = 0.084, G = 4.086], time: 21.72sec Epoch: 092/200, Loss: [D_real = 0.057, D_fake = 0.113, G = 4.969], time: 22.01sec Epoch: 093/200, Loss: [D_real = 0.072, D_fake = 0.088, G = 4.737], time: 21.77sec Epoch: 094/200, Loss: [D_real = 0.150, D_fake = 0.236, G = 5.893], time: 21.77sec Epoch: 095/200, Loss: [D_real = 0.121, D_fake = 0.123, G = 4.178], time: 21.93sec Epoch: 096/200, Loss: [D_real = 0.101, D_fake = 0.168, G = 5.016], time: 21.69sec Epoch: 097/200, Loss: [D_real = 0.040, D_fake = 0.085, G = 5.418], time: 21.55sec Epoch: 098/200, Loss: [D_real = 0.153, D_fake = 0.104, G = 4.984], time: 21.68sec Epoch: 099/200, Loss: [D_real = 0.542, D_fake = 0.544, G = 4.861], time: 21.97sec Epoch: 100/200, Loss: [D_real = 0.229, D_fake = 0.121, G = 4.621], time: 21.74sec > Accuracy at epoch 100 [real: 97%, fake: 99%]
Epoch: 101/200, Loss: [D_real = 0.062, D_fake = 0.206, G = 5.791], time: 21.98sec Epoch: 102/200, Loss: [D_real = 0.125, D_fake = 0.300, G = 6.427], time: 21.64sec Epoch: 103/200, Loss: [D_real = 0.158, D_fake = 0.122, G = 5.495], time: 22.11sec Epoch: 104/200, Loss: [D_real = 0.112, D_fake = 0.068, G = 4.546], time: 21.76sec Epoch: 105/200, Loss: [D_real = 0.066, D_fake = 0.270, G = 6.790], time: 22.06sec Epoch: 106/200, Loss: [D_real = 0.091, D_fake = 0.057, G = 4.560], time: 21.59sec Epoch: 107/200, Loss: [D_real = 0.089, D_fake = 0.043, G = 4.819], time: 21.61sec Epoch: 108/200, Loss: [D_real = 0.056, D_fake = 0.127, G = 5.026], time: 21.92sec Epoch: 109/200, Loss: [D_real = 0.178, D_fake = 0.272, G = 3.843], time: 21.68sec Epoch: 110/200, Loss: [D_real = 0.091, D_fake = 0.255, G = 5.615], time: 21.82sec Epoch: 111/200, Loss: [D_real = 0.122, D_fake = 0.122, G = 4.275], time: 21.84sec Epoch: 112/200, Loss: [D_real = 0.100, D_fake = 0.180, G = 5.019], time: 21.85sec Epoch: 113/200, Loss: [D_real = 0.123, D_fake = 0.062, G = 4.300], time: 21.77sec Epoch: 114/200, Loss: [D_real = 0.086, D_fake = 0.156, G = 5.417], time: 21.68sec Epoch: 115/200, Loss: [D_real = 0.137, D_fake = 0.064, G = 4.444], time: 21.69sec Epoch: 116/200, Loss: [D_real = 0.051, D_fake = 0.045, G = 4.531], time: 21.58sec Epoch: 117/200, Loss: [D_real = 0.054, D_fake = 0.056, G = 4.799], time: 21.59sec Epoch: 118/200, Loss: [D_real = 0.045, D_fake = 0.171, G = 5.375], time: 21.72sec Epoch: 119/200, Loss: [D_real = 0.114, D_fake = 0.193, G = 6.558], time: 21.73sec Epoch: 120/200, Loss: [D_real = 0.104, D_fake = 0.163, G = 5.800], time: 21.63sec > Accuracy at epoch 120 [real: 98%, fake: 100%]
Epoch: 121/200, Loss: [D_real = 0.153, D_fake = 0.104, G = 4.816], time: 21.69sec Epoch: 122/200, Loss: [D_real = 0.060, D_fake = 0.117, G = 6.667], time: 21.77sec Epoch: 123/200, Loss: [D_real = 0.042, D_fake = 0.136, G = 5.419], time: 21.79sec Epoch: 124/200, Loss: [D_real = 0.079, D_fake = 0.266, G = 5.975], time: 21.73sec Epoch: 125/200, Loss: [D_real = 0.061, D_fake = 0.218, G = 5.864], time: 21.68sec Epoch: 126/200, Loss: [D_real = 0.817, D_fake = 0.245, G = 3.185], time: 21.74sec Epoch: 127/200, Loss: [D_real = 0.473, D_fake = 0.196, G = 3.237], time: 21.71sec Epoch: 128/200, Loss: [D_real = 0.098, D_fake = 0.237, G = 4.490], time: 21.68sec Epoch: 129/200, Loss: [D_real = 0.201, D_fake = 0.095, G = 4.469], time: 21.89sec Epoch: 130/200, Loss: [D_real = 0.106, D_fake = 0.130, G = 4.117], time: 21.62sec Epoch: 131/200, Loss: [D_real = 0.320, D_fake = 0.078, G = 4.160], time: 21.69sec Epoch: 132/200, Loss: [D_real = 0.254, D_fake = 0.063, G = 3.754], time: 21.7sec Epoch: 133/200, Loss: [D_real = 0.084, D_fake = 0.113, G = 5.459], time: 21.62sec Epoch: 134/200, Loss: [D_real = 0.191, D_fake = 0.129, G = 4.404], time: 21.66sec Epoch: 135/200, Loss: [D_real = 0.144, D_fake = 0.024, G = 4.830], time: 21.84sec Epoch: 136/200, Loss: [D_real = 0.119, D_fake = 0.048, G = 5.230], time: 21.72sec Epoch: 137/200, Loss: [D_real = 0.124, D_fake = 0.020, G = 5.288], time: 21.71sec Epoch: 138/200, Loss: [D_real = 0.084, D_fake = 0.134, G = 4.797], time: 21.74sec Epoch: 139/200, Loss: [D_real = 0.072, D_fake = 0.088, G = 5.162], time: 22.06sec Epoch: 140/200, Loss: [D_real = 0.165, D_fake = 0.044, G = 4.763], time: 21.63sec > Accuracy at epoch 140 [real: 88%, fake: 98%]
Epoch: 141/200, Loss: [D_real = 0.082, D_fake = 0.036, G = 5.130], time: 21.68sec Epoch: 142/200, Loss: [D_real = 0.066, D_fake = 0.067, G = 4.151], time: 22.07sec Epoch: 143/200, Loss: [D_real = 0.182, D_fake = 0.118, G = 4.076], time: 22.08sec Epoch: 144/200, Loss: [D_real = 0.169, D_fake = 0.283, G = 4.874], time: 21.72sec Epoch: 145/200, Loss: [D_real = 0.069, D_fake = 0.167, G = 5.635], time: 21.81sec Epoch: 146/200, Loss: [D_real = 0.294, D_fake = 0.204, G = 4.635], time: 21.85sec Epoch: 147/200, Loss: [D_real = 0.151, D_fake = 0.061, G = 4.468], time: 21.56sec Epoch: 148/200, Loss: [D_real = 0.088, D_fake = 0.099, G = 4.917], time: 21.69sec Epoch: 149/200, Loss: [D_real = 0.213, D_fake = 0.051, G = 3.918], time: 21.53sec Epoch: 150/200, Loss: [D_real = 0.083, D_fake = 0.083, G = 4.674], time: 22.24sec Epoch: 151/200, Loss: [D_real = 0.085, D_fake = 0.033, G = 4.391], time: 22.02sec Epoch: 152/200, Loss: [D_real = 0.174, D_fake = 0.269, G = 4.916], time: 21.78sec Epoch: 153/200, Loss: [D_real = 0.206, D_fake = 0.219, G = 3.581], time: 21.64sec Epoch: 154/200, Loss: [D_real = 0.111, D_fake = 0.098, G = 4.278], time: 21.74sec Epoch: 155/200, Loss: [D_real = 0.241, D_fake = 0.232, G = 4.545], time: 21.71sec Epoch: 156/200, Loss: [D_real = 0.051, D_fake = 0.053, G = 4.909], time: 21.7sec Epoch: 157/200, Loss: [D_real = 0.063, D_fake = 0.133, G = 5.140], time: 21.97sec Epoch: 158/200, Loss: [D_real = 0.025, D_fake = 0.116, G = 5.342], time: 21.7sec Epoch: 159/200, Loss: [D_real = 0.068, D_fake = 0.113, G = 5.210], time: 21.75sec Epoch: 160/200, Loss: [D_real = 0.014, D_fake = 0.047, G = 5.497], time: 21.91sec > Accuracy at epoch 160 [real: 94%, fake: 100%]
Epoch: 161/200, Loss: [D_real = 0.058, D_fake = 0.104, G = 4.224], time: 21.78sec Epoch: 162/200, Loss: [D_real = 0.159, D_fake = 0.137, G = 5.223], time: 21.91sec Epoch: 163/200, Loss: [D_real = 0.073, D_fake = 0.121, G = 5.634], time: 21.71sec Epoch: 164/200, Loss: [D_real = 0.075, D_fake = 0.415, G = 5.290], time: 21.7sec Epoch: 165/200, Loss: [D_real = 0.169, D_fake = 0.083, G = 5.085], time: 21.75sec Epoch: 166/200, Loss: [D_real = 0.034, D_fake = 0.204, G = 6.189], time: 21.71sec Epoch: 167/200, Loss: [D_real = 0.092, D_fake = 0.177, G = 5.232], time: 22.03sec Epoch: 168/200, Loss: [D_real = 0.264, D_fake = 0.131, G = 4.127], time: 21.77sec Epoch: 169/200, Loss: [D_real = 0.152, D_fake = 0.113, G = 4.611], time: 21.74sec Epoch: 170/200, Loss: [D_real = 0.153, D_fake = 0.138, G = 4.498], time: 21.74sec Epoch: 171/200, Loss: [D_real = 0.099, D_fake = 0.065, G = 3.872], time: 21.71sec Epoch: 172/200, Loss: [D_real = 0.093, D_fake = 0.095, G = 5.010], time: 21.97sec Epoch: 173/200, Loss: [D_real = 0.097, D_fake = 0.163, G = 5.546], time: 21.92sec Epoch: 174/200, Loss: [D_real = 0.055, D_fake = 0.057, G = 5.143], time: 21.82sec Epoch: 175/200, Loss: [D_real = 0.058, D_fake = 0.073, G = 5.722], time: 21.76sec Epoch: 176/200, Loss: [D_real = 0.026, D_fake = 0.076, G = 5.435], time: 21.94sec Epoch: 177/200, Loss: [D_real = 0.066, D_fake = 0.022, G = 5.124], time: 21.83sec Epoch: 178/200, Loss: [D_real = 0.122, D_fake = 0.265, G = 6.568], time: 21.87sec Epoch: 179/200, Loss: [D_real = 0.245, D_fake = 0.209, G = 4.770], time: 21.77sec Epoch: 180/200, Loss: [D_real = 0.153, D_fake = 0.108, G = 4.339], time: 21.72sec > Accuracy at epoch 180 [real: 100%, fake: 97%]
Epoch: 181/200, Loss: [D_real = 0.108, D_fake = 0.117, G = 4.633], time: 21.97sec Epoch: 182/200, Loss: [D_real = 0.034, D_fake = 0.053, G = 4.950], time: 21.73sec Epoch: 183/200, Loss: [D_real = 0.097, D_fake = 0.205, G = 5.439], time: 22.42sec Epoch: 184/200, Loss: [D_real = 0.050, D_fake = 0.044, G = 4.637], time: 21.99sec Epoch: 185/200, Loss: [D_real = 0.156, D_fake = 0.041, G = 4.875], time: 22.42sec Epoch: 186/200, Loss: [D_real = 0.070, D_fake = 0.045, G = 5.375], time: 21.75sec Epoch: 187/200, Loss: [D_real = 0.064, D_fake = 0.249, G = 7.319], time: 21.88sec Epoch: 188/200, Loss: [D_real = 0.074, D_fake = 0.103, G = 5.076], time: 22.0sec Epoch: 189/200, Loss: [D_real = 0.062, D_fake = 0.042, G = 5.400], time: 21.75sec Epoch: 190/200, Loss: [D_real = 0.026, D_fake = 0.127, G = 7.020], time: 21.73sec Epoch: 191/200, Loss: [D_real = 0.019, D_fake = 0.172, G = 5.823], time: 21.66sec Epoch: 192/200, Loss: [D_real = 0.027, D_fake = 0.083, G = 6.206], time: 21.68sec Epoch: 193/200, Loss: [D_real = 0.110, D_fake = 0.089, G = 6.102], time: 21.73sec Epoch: 194/200, Loss: [D_real = 0.044, D_fake = 0.075, G = 5.475], time: 21.84sec Epoch: 195/200, Loss: [D_real = 2.356, D_fake = 63.535, G = 77.418], time: 21.73sec Epoch: 196/200, Loss: [D_real = 0.099, D_fake = 0.318, G = 2.791], time: 21.73sec Epoch: 197/200, Loss: [D_real = 0.085, D_fake = 0.165, G = 3.645], time: 21.71sec Epoch: 198/200, Loss: [D_real = 0.621, D_fake = 0.032, G = 4.384], time: 21.73sec Epoch: 199/200, Loss: [D_real = 0.492, D_fake = 0.340, G = 3.108], time: 21.66sec Epoch: 200/200, Loss: [D_real = 0.367, D_fake = 0.796, G = 3.248], time: 21.71sec > Accuracy at epoch 200 [real: 77%, fake: 100%]
Total time for training 200 epochs is 1h 12min 57.59sec sec
model.save('cnn_model.h5')
Generating and saving deepfakes using trained generator from dcgan
def save_fake_samples(g_model, latent_dim, n_samples=500):
x_fake, y_fake = generate_fake_samples(g_model, latent_dim, n_samples)
for i in range(len(x_fake)):
plt.axis('off')
plt.imshow(x_fake[i])
plt.savefig('/kaggle/working/FakeSamples/image{:04d}.png'.format(i+1))
os.makedirs('/kaggle/working/FakeSamples')
save_fake_samples(generator, latent_dim)
df_ims_ = []
for dirname, _, filenames in os.walk('/kaggle/working/FakeSamples'):
df_ims_.append(filenames)
# print(dirname, _, filenames)
df_ims = df_ims_[0]
# /kaggle/input/chest-ct-deepfakes/image0001.png
df_ims = ['/kaggle/working/FakeSamples/'+i for i in df_ims]
'/kaggle/working/FakeSamples/image0379.png'
temp = cv.imread(df_ims[0])
<class 'numpy.ndarray'>
def preprocessDeepFakes(imgs_inp):
preprocessed = []
for im in imgs_inp:
im = cv.imread(im)
if type(im) == np.ndarray:
im = cv.cvtColor(im, cv.COLOR_BGR2GRAY)
preprocessed.append(cv.resize(im, (32, 32)))
return preprocessed
deepfakespre = preprocessDeepFakes(df_ims)
deepfakespre = np.array(deepfakespre)/255
deepfakes = deepfakespre.reshape((deepfakespre.shape[0], deepfakespre.shape[1], deepfakespre.shape[2], 1))
y_fake = np.random.randint(low=0, high=2, size=(500, 1))
deepfakes.shape, y_fake.shape
((500, 32, 32, 1), (500, 1))
model.evaluate(deepfakes, y_fake)
16/16 [==============================] - 0s 3ms/step - loss: 7.2324 - accuracy: 0.5000
[7.232410907745361, 0.5]
As seen above, the accuracy of the model significantly decreases on a deepfakes dataset, to only 50%